Single-set document archive intelligent reorganization method based on AI agent
By employing a fully automated compilation solution based on seven AI-powered engines and the BERT model, the problems of low efficiency, inconsistent standards, and security in single-set electronic document archiving have been solved, achieving efficient, secure, and compliant electronic document management.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- FUJIAN YIRONG INFORMATION TECH
- Filing Date
- 2026-04-21
- Publication Date
- 2026-07-21
AI Technical Summary
Existing technologies lack the ability to intelligently compile documents throughout their entire lifecycle in single-set electronic document archiving, resulting in low efficiency, inconsistent standards, susceptibility to errors, and difficulty in meeting security and compliance requirements.
A fully automated compilation solution is constructed using seven AI-based intelligent agents (metadata extraction, classification, security classification, retention period determination, document number generation, and boxing, etc.). Combined with the BERT model and the knowledge graph of security points, it realizes the automatic classification, security classification, document number generation, and boxing of electronic documents.
It achieves fully automated operation, improving efficiency, unifying standards, ensuring safety and compliance, meeting high requirements, and being low-carbon and energy-saving, in line with the national dual-carbon strategy.
Smart Images
Figure CN122432262A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of archival management technology, and in particular to an intelligent compilation method for single-set document archives based on AI intelligent agents. Background Technology
[0002] With the advancement of the "single-set" archiving reform, electronic documents no longer generate paper copies; they are archived directly in electronic form as the sole original. Traditional manual compilation methods are inefficient, inconsistent, and prone to errors, and heavily rely on the experience of archivists. Existing technologies mostly focus on OCR recognition or metadata extraction, lacking integrated intelligent compilation capabilities for the single-set system, encompassing "classification-confidentiality designation-duration period-file number-boxing-archiving," making it difficult to meet the high requirements of the "Electronic Document Archiving Specification" (DA / T 92-2022) for authenticity, completeness, usability, and security. Summary of the Invention
[0003] To address the aforementioned issues, the present invention aims to provide an intelligent compilation method for single-set document archives based on AI intelligent agents, thereby achieving automated, intelligent, and compliant compilation of electronic documents throughout their entire lifecycle and resolving problems such as excessive manual intervention, inconsistent standards, and difficulty in controlling risks.
[0004] To achieve the above objectives, the present invention adopts the following technical solution: A method for intelligent compilation of single-set document archives based on AI intelligent agents includes the following steps: Step 10: Input the electronic document into the metadata intelligent extraction engine. The metadata intelligent extraction engine includes a rule engine and a CRF model. The rule engine performs a preliminary screening of the electronic document, and then the preliminary screening results are input into the CRF model to output metadata. Step 20: Input the metadata into the classification engine. The classification engine is based on the BERT model and combines a neighborhood vocabulary to classify electronic documents and output a classification number. Step 30: Input the metadata into the security classification engine, which combines the security knowledge graph to determine the security level of the electronic document and mark the security points; Step 40: Input the metadata into the retention period determination engine. The retention period determination engine combines the pre-built ternary decision table and generates a retention period suggestion through metadata mapping and matching. Step 50: Input the metadata into the file number generation engine, and output the file number according to the preset file number rules; Step 60: Input the metadata into the automatic boxing engine, match the metadata with the preset boxing strategy, box the electronic documents into groups, automatically assign box numbers, and generate a box number file directory.
[0005] The present invention has the following beneficial effects: 1. Fully automated process: The number of manual intervention points has been reduced to two (confidentiality verification and archiving confirmation), significantly improving efficiency; 2. Standardized and Unified: Based on standardized training, the classification, expiration date, document number, and packaging rules are 100% consistent; 3. Security and Compliance: Meets the Level 3 requirements of the "Electronic Document Archiving Standard" and the "Security Level Protection Standard for Archives Information Systems"; 4. Low carbon and energy saving: Eliminating paper copies saves more than 95% of paper annually, which is in line with the national dual carbon strategy. Attached Figure Description
[0006] Figure 1 This is a schematic diagram of the overall architecture of the present invention; Figure 2 This is a diagram illustrating the construction process of the intelligent agent of the present invention; Figure 3 This is a flowchart illustrating the implementation of the present invention; Figure 4 This is a flowchart illustrating the workflow of the metadata intelligent extraction engine of the present invention. Figure 5 This is a flowchart of the classification engine of the present invention; Figure 6 This is a schematic diagram of the knowledge graph of dense points in this invention. Detailed Implementation
[0007] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments: This embodiment proposes an intelligent compilation solution for single-set document archives based on AI intelligent agents, consisting of "seven engines + one certificate", with the overall architecture as follows: Figure 1 As shown, an intelligent document compilation system is built through a workflow approach to achieve automated, intelligent, and compliant compilation of electronic documents throughout their entire lifecycle.
[0008] See Figures 1 to 3 A method for intelligent compilation of single-set document archives based on AI intelligent agents includes the following steps: After the user initiates the reorganization request, steps 10 to 40 are executed; Step 10, please refer to Figure 4 Metadata is extracted from electronic documents using a metadata intelligent extraction engine. This engine comprises a rule engine and a CRF model, and the execution steps are as follows: Step 11: Define regular expressions for extracting metadata and build a metadata extraction rule base; Step 12: Input the formatted electronic document, parse the file (using the open-source Apache PDFBox tool), and extract the document text information; Step 13: The rule engine combines with the metadata extraction rule library to conduct a preliminary screening on the official document text information, and obtains a preliminary metadata extraction result.
[0009] For example: Document number rule "^〔\d{4}〕\d+号 ", and the output result is "〔2025〕12号".
[0010] Date of issue rule "\d{4}年\d{1,2}月\d{1,2}日", and the output result is "2025年12月02日".
[0011] Step 14: Input the preliminary metadata extraction result and the official document text information into the CRF model, and perform sequence labeling through the CRF model to complete the accurate extraction of metadata. The CRF model is a discriminative probability model for sequence labeling tasks. It can label the dependency relationship between the preliminary metadata extraction result and adjacent elements, solve boundary ambiguity, and achieve the accurate extraction of metadata.
[0012] For example: Document number "闽\n电发〔2025〕12号" with line break across lines; Title粘连正文 "关于XXX的通知各相关单位:".
[0013] Step 15: Post-correction processing: Convert non-uniform metadata into a standard data format.
[0014] For example: Date "2025年12月02日" is uniformly converted to "20251202"; Document type is automatically classified as "通知 / 报告 / 请示 / 函", etc. Extract the filing department and fonds number from the department and fonds information of the current operating user. (The fonds number is the number of the fonds to which the current user belongs, obtained through the built-in fonds table. The fonds table contains information such as fonds number and fonds name. Each fonds is associated with the user unit on a one-to-one basis and is written together when generating metadata.) Output: Metadata such as the filing department, fonds number, title, document type, document type, document number, responsible unit, year, date of issue, document content, etc. of the official document.
[0015] Step 20: Please refer to Figure 5 , and classify the electronic official document through the classification engine. The specific steps are as follows: Pre-defined classification tree: The files are divided into several categories, and each category corresponds to a classification tree. The classification tree adopts a three-level structure: function-business-matter; each function, business, and matter is a classification node, and each classification node is numbered. For example: Pre-defined 0-5 major category classification tree: 0 Party and mass work, 1 administrative management, 2 business management, 3 production technology management, 4 financial audit, 5 labor and personnel; the corresponding numbering method of the Party and mass work classification tree is: Party and mass work (0) - Party affairs work (01) - Party committee meeting minutes (011).
[0016] A pre-defined thesaurus of official documents and classification indexing rules for archives are used. For example, a pre-defined thesaurus of official documents of Party and government organs and classification indexing rules for archives are used. Then, a domain thesaurus is constructed that associates the thesaurus rules with the classification numbers. The construction process of the domain thesaurus is as follows: a triple (thesaurus, classification number, relation type) is constructed from the thesaurus of official documents, the classification indexing rules for archives, and the third-level item nodes in the classification tree. The classification number is determined by querying the classification tree based on the item name. The domain thesaurus consists of multiple triples. The relation types include equivalence, subordination, and relevance. For example: equivalence relation: the thesaurus "fiscal budget" is equivalent to the classification "fiscal management"; subordination relation: the thesaurus "fiscal budget execution status" is a sub-part of "fiscal budget" and can be classified under the "fiscal management" classification; relevance relation: the thesaurus "fiscal audit" and "fiscal budget" have a certain correlation, but are not completely equivalent or have a partial relationship. Each relation type is represented by a different value, for example, {"equivalence" = 0, "subordination" = 1, "relevance" = 2}.
[0017] This invention employs a hybrid model combining BERT, a domain-specific thesaurus, and reinforcement learning to automatically match the aforementioned official document subject thesaurus and archival classification indexing rules, calculating classification information. The specific steps are as follows: Metadata, including title, document content, document type, and document number, is input into the BERT model. Semantic encoding is performed to generate multi-dimensional embedding vectors, which can be word vectors or text vectors. When the input information is a word, a word vector is output; when the input information is a sentence, a sentence vector is output; and when the input vector is text, a text vector is output. For example, a 768-dimensional embedding vector is generated. The BERT model can generate corresponding word vectors and text vectors based on the input words and text. The domain vocabulary embedding vector is generated through TransE, and the process is as follows: a.) Data Preparation Positive sampling: triplet data Negative sampling: For each triplet, 5 incorrect classification results are randomly generated as negative samples; b.) Initialization Initialize the triplet (topic term, classification number, relation) embedding vector using Xavier, with a dimension of 128; c.) Scoring function The scores of the triples are calculated using the scoring function of the TransE model; d.) Loss function The model is optimized using the Margin Loss loss function so that the score of the triple is less than the score of the negative sample. e.) Optimizer The optimization algorithm (Adam) is used to learn and update the embedding vectors, with a learning rate set to 2e. -5 ; f.) Training Cycle After 20 training rounds, a 128-dimensional domain vocabulary embedding vector is generated.
[0018] Attention is fused between the multidimensional embedding vector and the domain vocabulary embedding vector, and the weights of the multidimensional embedding vector are calculated using a normalized exponential function; for example, the multidimensional embedding vector output by the BERT model is a text vector H. bert ∈R 768 The domain vocabulary embedding vector is H reg ∈R 128 The trainable domain vocabulary matrix W∈R (768×128) As the weight matrix, the formula for calculating the cross-attention score is: , The attention weight represents the degree of dependence of the text vector on the domain vocabulary embedding vector. The calculation process of the formula is as follows: 1) Matrix multiplication: Multiply the sentence vector output by BERT with the weight matrix to obtain an intermediate vector; 2) ) 1) Multiply the intermediate vector by the transpose of the domain vocabulary embedding vector (T represents the matrix transpose operation) to obtain a scalar value; 2) Normalize the above scalar value using the Softmax function to obtain the attention weight, which represents the degree of dependence of the text vector on the domain vocabulary vector. Output the classification number, which is based on the attention weight calculated by the Softmax function, prioritizing the highest weight.
[0019] The process of generating the trainable domain vocabulary matrix W is as follows: a.) Data Preparation Sample data (metadata that has been classified) Domain vocabulary embedding vector b.) Initialization Initialize a trainable domain vocabulary matrix W using Xavier, with dimensions of 768×128, where 768 is the dimension of the BERT output vector and 128 is the dimension of the domain vocabulary embedding vector.
[0020] c.) Cross-attention score calculation Cross-attention formula calculates sample data scores d.) Loss function The classification loss is calculated using the binary cross-entropy loss function. e.) Optimizer The optimizer (Adam) is used to learn and update the matrix W, with a learning rate set to 2e. -5 ; f.) Training Cycle Based on the amount of sample data, iterative training is performed to generate a trainable domain vocabulary matrix W.
[0021] The domain vocabulary is subjected to reinforcement learning, and its iterative steps are as follows: During the user's metadata review process, each time the classification number is confirmed or modified, the system automatically records (text, correct classification number); each time new data is added, the training cycle is fine-tuned using the new data, with a learning rate of 2e. -5 It enables online updates of the domain terminology, allowing users to learn as they use it.
[0022] Step 30: Determine the security classification of the electronic document using the security classification engine. The specific steps are as follows: Please see Figure 6Based on relevant national and industry regulations and combined with multi-source data such as classified terms and sensitive words, a four-layer knowledge graph of confidential information is constructed, including a regulation layer, an industry layer, a matter layer, and a confidential information layer. The regulation layer contains national and industry confidentiality regulations and policy documents. For example, the *National Secrets Law* stipulates the scope, classification, and confidentiality period of state secrets, while the *Electric Power Industry Confidentiality Management Measures* stipulates confidentiality management in the electric power industry. The industry layer includes detailed confidentiality regulations for different applications in different industry sectors. For example, the *Electric Power Industry Confidentiality Management Measures* stipulates confidentiality requirements for power data, customer data, and power grid operation data. The matter layer further refines the information to specific business matters, such as: power grid operation data management (involving peak load, power generation, transmission losses, etc.); financial audit report management (involving sensitive information such as financial data and audit conclusions); and customer electricity information management (involving customer electricity data and electricity bills). The confidential information layer contains specific information that needs to be kept confidential, i.e., specific data or text fragments that need to be kept confidential. For example: power grid operation data, the peak load of a certain region's power grid is "12000MW", classified as "secret" with a confidentiality period of "permanent"; customer electricity consumption data, a customer's monthly electricity consumption is "5000kWh", classified as "internal" with a confidentiality period of "10 years"; audit conclusion: a certain power company has violated regulations, classified as "confidential" with a confidentiality period of "30 years".
[0023] Using a dense-point sensitive feature extraction method, the system outputs a security classification suggestion, including: security classification, confidentiality period, and can also output the regulations on which it is based. Rule-based matching: Based on the content of the input electronic document, the industry to which the text belongs is first identified, and then the applicable regulatory documents are matched. Industry-level refinement: Matching electronic documents with industry levels to identify the application type of the electronic document; Item-level positioning: Matching electronic documents with industry-level data to identify specific business items; Secret point layer processing: Match the identified specific business items with the secret point layer to determine the specific data or text fragments that need to be kept confidential, mark the secret points and label the security level; determine the security level of the document according to the principle of prioritizing the highest security level.
[0024] For example: Input: For example, the content of a document containing power grid operation data, customer electricity consumption data, and audit conclusions; Regulatory matching: Based on the content of the input document, the system first identifies that the text belongs to the power industry and matches it with the "National Secrecy Law" and the "Power Industry Secrecy Management Measures". Industry-level refinement: Further refinement to the management of financial audit reports and customer electricity consumption information in the power industry, and identify the confidentiality requirements related to the management of financial audit reports; Item-level positioning: Under the financial audit report management items, identify specific key points, such as power grid operation data "1200MW" and customer electricity consumption data "5000kWh".
[0025] Classified information processing: Identified classified information is marked with a classification level of "Secret" or "Internal" and a confidentiality period of "Permanent" or "10 Years". Documents are classified as "Secret" with a confidentiality period of "Permanent" based on the highest classification level, and the relevant provisions of the "Power Industry Confidentiality Management Measures" are provided.
[0026] Step 40: Storage Period Determination Engine; The system pre-sets relevant documents regarding retention periods and constructs a ternary decision table and a matter keyword rule table. The ternary decision table and the matter keyword rule table are linked through matters. The ternary decision table consists of functions, matters, and retention periods, while the matter keyword rule table consists of matters and rule expressions, supporting customization (the rule expression structure is: "(metadata name) logical word [keyword]", for example: "(title) contains [Party conduct and integrity, democratic life meeting]").
[0027] A pre-set table of the correspondence between archiving departments and functions is provided. For example, if the archiving department is the Organization Department, the corresponding functions are the organization structure and personnel management, cadre selection and appointment, performance evaluation, salary and performance, and other related functions.
[0028] Input: Title, archiving department, classification number, and document content; The steps for determining the storage period are as follows: Based on the archiving department, determine the functions using the comparison table between archiving departments and their functions; Input the function, and output multiple first-choice alternatives through a three-element decision table; Determine the second alternative item corresponding to the classification number based on the classification number and the classification tree; The second alternative item is matched with the first alternative item. If they match, the second alternative item is selected as the confirmed item. If they do not match, the corresponding rule is extracted from the item keyword rule table based on the first alternative item. The rule is then matched with the metadata to determine the item. Based on the defined functions and tasks, the retention period is determined by matching the aforementioned ternary decision table.
[0029] Output: Retention period.
[0030] After the electronic document compilation is completed and the user initiates the archiving instruction, proceed with steps 50 to 70.
[0031] Step 50: Output file number via file number generation engine: A pre-defined five-segment file numbering rule is used, including record number, year, classification number, retention period, and item number. Standard file numbers are generated through this rule. The specific steps are as follows: Input: metadata record number, year, classification number, and retention period; Part number generation: The part number is automatically generated by incrementing according to the first four segments of metadata in the file number rules, namely the record number, year, classification number and retention period; File number generation: Based on the five-segment file number rule, extract the corresponding record number, year, classification number, retention period and item number, and concatenate them to generate the file number; Output: File number.
[0032] Step 60: Complete file boxing using the automatic boxing engine; A three-tiered boxing strategy with pre-set year, classification number, and shelf life, and pre-set boxing capacity specifications; Retrieve the metadata of the file with the generated file number; The boxes are sorted and boxed according to the boxing strategy, and box numbers are automatically assigned, generating a box number file directory; Output: Box number, box spine information, and box contents XML.
[0033] Step 70: Detect electronic documents using the archiving security engine; Authenticity: Extract the electronic signature of the electronic document, perform CA verification, and determine the authenticity of the electronic document; Integrity: Write the metadata into XML, then merge and compress the XML and the original file into a ZIP file, calculate the hash value of the file, obtain the hash value of the archived data package, and record the archive timestamp; Availability: Using the Apache PDFBox open-source tool, files are uniformly converted to PDF / A format, ensuring long-term file availability; Security: The system's antivirus, encryption, and auditing measures.
[0034] Strengthening of national cryptographic SM2 / SM3 / SM4 algorithms and hardware security module HSM.
[0035] Each electronic document is generated with a unique identification code after being archived, including: file number, box number, archived data packet hash, and archived timestamp, supporting full lifecycle tracking of subsequent borrowing, transfer, and destruction.
[0036] The above description is merely a specific embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural transformations made based on the content of the present invention specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.
Claims
1. A method for intelligent compilation of single-set document archives based on AI intelligent agents, characterized in that: Includes the following steps: Step 10: Input the electronic document into the metadata intelligent extraction engine. The metadata intelligent extraction engine includes a rule engine and a CRF model. The rule engine performs a preliminary screening of the electronic document, and then the preliminary screening results are input into the CRF model to output metadata. Step 20: Input the metadata into the classification engine. The classification engine, based on the output of the BERT model, combines the neighborhood vocabulary and then uses the cross-attention fusion module to classify the electronic documents and output the classification number. Step 30: Input the metadata into the security classification engine, which combines the security knowledge graph to determine the security level of the electronic document and mark the security points; Step 40: Input the metadata into the retention period determination engine. The retention period determination engine combines the pre-built ternary decision table and generates the retention period through metadata mapping and matching. Step 50: Input the metadata into the file number generation engine, and output the file number according to the preset file number rules; Step 60: Input the metadata into the automatic boxing engine, match the metadata with the preset boxing strategy, box the electronic documents into groups, automatically assign box numbers, and generate a box number file directory.
2. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 1, characterized in that: The specific execution process of step 10 is as follows: Define a regular expression for extracting metadata and construct a metadata extraction rule base; the rule engine, in conjunction with the metadata extraction rule base, performs a preliminary screening of the text information of electronic documents to obtain preliminary metadata extraction results; input the preliminary metadata extraction results and text information into the CRF model, perform sequence labeling through the CRF model, use context features and sequence dependencies for text labeling, complete the extraction of metadata, perform post-processing correction on the extracted metadata, and output the metadata; the metadata includes the title, document type, document name, document number, responsible unit, date of issuance, document content, and / or archiving department of the document.
3. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 2, characterized in that: The specific execution process of step 20 is as follows: Pre-defined classification tree: The files are divided into several categories, and each category corresponds to a classification tree. The classification tree adopts a three-level structure: function-business-item; each function, business and item is a classification node, and each classification node is numbered. A pre-set thesaurus of official document subject terms and archival classification indexing rules are used to construct a domain thesaurus that associates the subject term rules with classification numbers. The domain thesaurus construction process involves: constructing triples (subject term, classification number, relation type) from the official document subject terms, archival classification indexing rules, and items in the classification tree; multiple triples constitute the domain thesaurus; determining the classification number by querying the classification tree based on the item name; the relation types include equivalence, genus, and relevance, each represented by a different numerical value; and automatically matching the relevant official document subject terms and archival classification indexing rules using a BERT model to calculate classification information. The specific steps are as follows: The metadata is input into the BERT model for semantic encoding, generating multi-dimensional embedding vectors; the BERT model can generate corresponding word vectors and text vectors based on the input words and text. The domain vocabulary is used to generate a domain vocabulary embedding vector through the TransE model; Attention is fused between multidimensional embedding vectors and domain vocabulary embedding vectors, and data weights are calculated using a normalized exponential function. Specifically, the BERT model outputs a text vector H. bert ∈R 768 The domain vocabulary embedding vector is H reg ∈R 128 The trainable domain vocabulary matrix is W∈R (768×128) As the weight matrix, the formula for calculating the cross-attention score is: , The attention weight represents the degree to which the text vector depends on the domain vocabulary embedding vector; Based on the attention weight values, the result with the highest weight is selected first, and the classification number is output.
4. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 2, characterized in that: The specific execution process of step 30 is as follows: A four-layer knowledge graph of confidential information is constructed, including a regulation layer, an industry layer, a matter layer, and a confidential information layer. The regulation layer contains national and industry confidentiality regulations and policy documents. The industry layer includes detailed confidentiality regulations for different applications in different industry fields. The matter layer further refines the information to specific business matters based on the industry layer. The confidential information layer contains specific information that needs to be kept confidential. A sensitive feature extraction method at the dense point level is used to output a security classification suggestion, including the security classification, the confidentiality period, and the regulations on which it is based. Rule-based matching: Based on the content of the input electronic document, the industry to which the text belongs is first identified, and then the applicable regulatory documents are matched. Industry-level refinement: Matching electronic documents with industry levels to identify the application type of the electronic document; Item-level positioning: Matching electronic documents with industry-level data to identify specific business items; Secret Point Layer Processing: Match the identified specific business items with the secret point layer to determine the specific data or text fragments that need to be kept confidential, and mark the secret points with a confidentiality level.
5. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 2, characterized in that: The specific execution process of step 40 is as follows: The relevant documents with pre-defined retention periods are used to construct a ternary decision table and a matter keyword rule table; the ternary decision table and the matter keyword rule table are associated through matters; the ternary decision table consists of functions, matters, and retention periods, and the matter keyword rule table consists of matters and rule expressions. Pre-set archiving department and function correspondence table; Input: Title, archiving department, classification number, and document content; The steps for determining the storage period are as follows: Based on the archiving department, determine the functions using the comparison table between archiving departments and their functions; Input the function, and output multiple first-choice alternatives through a three-element decision table; Determine the second alternative item corresponding to the classification number based on the classification number and the classification tree; Match the second alternative with the first alternative. If they match, the second alternative is selected as the confirmed alternative. If there is a discrepancy, based on the first alternative item, the corresponding rule is extracted from the item keyword rule table, and the rule is matched with the metadata to determine the item; Based on the defined functions and tasks, match the aforementioned ternary decision table to determine the retention period; Output: Retention period.
6. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 2, characterized in that: Step 50 generates standard file numbers according to the file numbering rules. The specific execution process is as follows: Pre-defined five-segment file numbering rules: including record number, year, category, retention period, and item number; Input: Metadata; Part number generation: The part number is automatically generated by incrementing the first four segments of metadata, namely, the record number, year, category, and retention period, according to the file numbering rules. File number generation: Based on the five-segment file number rule, extract the corresponding record number, year, classification number, retention period and item number, and concatenate them to generate the file number; Output: File number.
7. The method for intelligent compilation of single-set document archives based on AI intelligent agents according to claim 2, characterized in that: The specific execution process of step 60 is as follows: A three-tiered boxing strategy with pre-defined year, category, and storage period, and pre-defined boxing capacity specifications; Retrieve the metadata of the file with the generated file number; The boxes are sorted and boxed according to the boxing strategy, and box numbers are automatically assigned, generating a box number file directory; Output: Box number, box spine information, and box contents XML.